On 2022 March 4, the object known as WE0913A crashed into the Moon after several close flybys of the Earth and
the Moon in the previous three months. Leading up to impact, the identity of the lunar impactor was up for debate,
with two possibilities: the Falcon 9 from the DSCOVR mission or the Long March 3C from the Chang’e 5-T1
mission. In this paper, we present a trajectory and spectroscopic analysis using ground-based telescope
observations to show conclusively that WE0913A is the Long March 3C rocket body (R/B) from the Chang’e
5-T1 mission. Analysis of photometric light curves collected before impact give a spin period of 185.221 ± 6.540 s
before the first close Earth flyby on 2022 January 20 and a period of 177.754 ± 0.779 s, both at a 1σ confidence
level, before the second close Earth flyby on 2022 February 8. Using Markov Chain Monte Carlo sampling and a
predictive light curve simulation based on an anisotropic Phong reflection model, we estimate both physical and
dynamical properties of the Chang’e 5-T1 R/B at the start of an observation epoch. The results from the Bayesian
analysis imply that there may have been additional mass on the front of the rocket body. Using our predicted
impact location, the Lunar Reconnaissance Orbiter was able to image the crater site approximately 7.5 km from the
prediction. Comparing the pre- and post-impact images of the location shows two distinct craters that were made,
supporting the hypothesis that there was additional mass on the rocket body
2. 2. Observations and Data Reduction
The data products used directly in this research are the time
history of reduced photometric brightness measurements (light
curves) taken from imagery data acquired with each telescope.
Astrometric position measurements are only used indirectly,
being needed for orbit determination to allow for later
trajectory propagation to support the reflective dynamics model
(see Sections 6.1.2 and 6.1.3). All the imagery data used were
collected from two telescopes located in Arizona, USA. Some
basic information for each of the sensors can be found in
Table 1.
All observations from Leo-20 were taken with an open/clear
filter, and those from RAPTORS I were taken using the Sloan
Digital Sky Survey g′ and r′ photometric filters with band
passes (>50% transmission) of 401–550 nm and 555–695 nm,
respectively. Visible spectra (450–950 nm) from RAPTORS I
were taken with a 30 line mm−1
transmission diffraction
grating installed in the light path, resulting in a 14.6 nm px−1
(R ∼ 30 at 450 nm) spectral resolution. Further details about the
gratingʼs throughput in each order and the process for
determining the spectral resolution can be found in Battle
et al. (2022, 2023). The observation geometry, dates, and
number of images taken can be found in Table 2.
The reduction and analysis of the telescope data are
accomplished by first calibrating all images with dark, bias,
and flat-field frames in order to reduce the amount of errant
noise and vignetting in the images caused by the sensor system.
Once the images are calibrated, the plate solution is calculated
using the GAIA Data Release 2 star catalog (Gaia
Collaboration 2018) and open-source software package Scamp
(Bertin 2006). After a successful plate solution is found, the
centroid and signal (flux) from the desired target (WE0913A)
may be extracted via aperture photometry, accomplished here
with the open-source software Source-Extractor (Bertin &
Arnouts 1996). Reported observations are flux calibrated with
up to a first-order correction calculated from solar analog stars
across all the frames in a data set. Solar analog stars are
selected based on the criteria given in Andrae et al. (2018).
Spectra are extracted from RAPTORS I images by summing
the flux in each column along the spectrum. Wavelength is
determined using the distance (in number of columns only)
from the centroid multiplied by the grating resolution. The flux
in each wavelength unit is divided by the flux in the same
wavelength unit for a solar analog spectrum to obtain a
reflectance value. The reflectance spectrum is then normalized
to unity at 700 nm.
3. WE0913A Trajectory Analysis
The Chang’e 5-T1 mission launched on 2014 October 23 at
18:00 UTC from the Xichang Satellite Launch Center, and
according to the Long March 3C user manual (Cen et al. 2011),
during a “typical flight sequence” the second third-stage
powered flight (which takes the payload from an Earth parking
orbit onto its translunar trajectory) occurs between 1323.2 and
1494.9 s post liftoff. Preceding this burn, the rocket/payload
combination is in a “coast phase” for roughly 11 minutes. This
point when the rocket/payload combination leaves its Earth
parking orbit marks the periapsis of its lunar transfer orbit.
Using observations collected in 2015 March, when
WE0913A was first observed by the Catalina Sky Survey, we
can fit an orbit and propagate the results backward in time to
observe the history of the trajectory of the object. A plot of the
trajectory of WE0913A is given in Figure 1.
Table 1
Relevant Parameters for the Sensors Used to Collect Data of the Chang’e 5-T1 R/B
Name Location Aperture (m) FOV (arcmin) Pixel Size (as px−1
)
Leo-20 Arizona, USA 0.52 81 × 81 2.4
RAPTORS I Arizona, USA 0.61 16 × 16 0.95
Note. FOV stands for Field of View.
Table 2
Summary of Observation Epochs and Geometry for the Collected Data
Name Date (UTC) Range (km) V Mag (Pred.) Phase Angle (°) No. of Obs.
Leo-20 2022/01/20 169,000 15.3 91.4 400
2022/02/08 144,000 15.2 102.5 300
RAPTORS I 2022/02/08 144,000 15.2 102.5 211
Figure 1. Example of the WE0913A orbit (black curve) from 2015 March 15
propagated backward until it hits its perigee on 2014 October 23. Orbits are
shown in the geocentric MEME J2000 equatorial reference frame. The Moon’s
orbit (gray curve) and Earth’s location (black dot) are included for reference.
The rocket body had close encounters with the moon on both 2015 February 13
and 2014 October 28.
2
The Planetary Science Journal, 4:217 (12pp), 2023 November Campbell et al.
3. In Figure 2, we show a short ground track of the backward
propagation up to perigee (denoted by a dot). The Xichang
launch facility is also shown for reference (marked as a star).
The time of perigee in Figures 1 and 2 should correspond with
the time window between coast and translunar orbit injection
that we described above, and we find that indeed it does. By
following the description in the Long March 3C user manual,
this phase should have occurred between 18:22:03.2 and
18:24:54.9 UTC on 2014 October 23. The results of our
backward propagation put this perigee at 18:24:34.3 UTC.
Moreover, the Long March 3C user manual lists the expected
sublatitudes and sublongitudes for the launch vehicle during its
initial phases of flight. Figure 3 shows the two launch phases as
large dots and our derived perigee as a small dot. Again, the
launch facility is included as a star for reference. As can be
seen in this figure, the perigee falls squarely between these two
launch phases, both geographically and in time.
The mean residual in our orbit determination from the
observations of WE0913A used to recreate this trajectory is
0 3. In this model, we estimate parameters corresponding to
solar radiation pressure (SRP), which must be considered for
this length of propagation. For instance, without taking into
account SRP, the time of perigee for this object is delayed to
20:55:30.8 UTC on 2014 October 23. A difference of 2.5 hr
and a perigee slightly off-track from the launch path! However,
the SRP parameters for a tumbling rocket body can be
extremely hard to estimate and are, in fact, changing in time.
As such, the parameters used represent a mean fit.
Based on this analysis, we show that the trajectory of
WE0913A matches the nominal mission plan published in the
Long March 3C user manual (Cen et al. 2011). Additionally,
the WE0913A trajectory strongly indicates a launch from the
Xichang Satellite Launch Center on 2014 October 23 at
approximately 1800 UTC. As such, we may conclude that this
object is likely to be the Long March 3C R/B (NORAD ID
40284, COSPAR ID 2014-065B) that launched the Chang’e
5-T1 mission.
4. Spectral Analysis
Spectroscopy has been shown to be a useful tool to identify
space objects in Earth orbit based on their reflectance properties
in the visible wavelengths (0.4–1.0 μm; Battle et al. 2023).
Building upon the trajectory analysis results, we have spectral
observations that favor the conclusion that the observed object
was indeed the Chang’e 5-T1 R/B instead of the Falcon 9 R/B
from the DSCOVR mission. In order to support this
conclusion, we needed to observe comparison objects that
had been launched at a similar time as the Chang’e 5-T1 R/B
so that the materials used and any alterations from space aging
would be similar (Jorgensen et al. 2001). Observations were
taken of a Falcon 9 R/B (NORAD ID 40108, COSPAR ID
2014-046B) launched in 2014 August and a Long March 3C
R/B (NORAD ID 43875, COSPAR ID 2018-110B) launched
in 2018 December. These will be used to compare to the
targets: the Falcon 9 R/B (NORAD ID 40391, COSPAR 2015-
007B) used to launch the DSCOVR mission in 2015 February
and the Long March 3C R/B (NORAD ID 40284, COSPAR
ID 2014-065B) used to launch the Chang’e 5-T1 in 2014
October.
Figure 4 shows the collected spectrum of the lunar impactor
(WE0913A) with spectra of the comparison objects and error
bars representing the standard deviation of the reflectance
Figure 2. Ground track of WE0913A (red curve) showing the orbit propagated
backward to perigee (red dot) on 2014 October 23 at 18:24:34.3 UTC
(24 minutes after Chang’e 5-T1 launch). Only a short time near perigee is
shown for simplicity. The Xichang Satellite Launch Center, from which the
Chang’e 5-T1 mission was launched, is shown as a red star.
Figure 3. Zoomed-in view of the WE0913A calculated perigee (red dot) with
the context of the coast and final burn stages of launch (large blue dots). The
Chang’e 5-T1 R/B translunar trajectory should have a perigee between these
two points, and we see that this is the case. The time window between the two
launch phases is T + 22.053 minutes and T + 24.915 minutes and our
calculated perigee is at T + 24.572 minutes. The Xichang Satellite Launch
Center is again shown as a red star.
Figure 4. Normalized reflectance spectra from RAPTORS I showing a
comparison between a Falcon 9 R/B (squares), a Long March 3C R/B
(triangles), and the lunar impactor (dots). A third-order polynomial fit to the
WE0913A data is provided as a reference. It is clear from the figure that the
lunar impactor is a much closer spectral match to the Long March 3C R/B than
the Falcon 9 R/B.
3
The Planetary Science Journal, 4:217 (12pp), 2023 November Campbell et al.
4. measured across multiple images taken during observations.
Since the WE0913A object was very faint for the RAPTORS I
telescope (V mag 15.2) during the observations, the resulting
low-resolution spectrum is noisy, so there is also included a
plot of a third-order polynomial fit to the data for reference.
The spectrum of the Falcon 9 R/B (denoted by squares in
Figure 4) is spectrally flat, reflecting blue (∼400 nm) and red
(∼900 nm) light almost equally. The spectra for both the
known Long March 3C R/B (triangles) and the lunar impactor
(dots) are red-sloped, reflecting more red light than blue light.
Although spectral slope is not necessarily diagnostic of what
material or object is being observed, we can use these results to
say which of the comparison objects the lunar impactor more
closely matches. Using the third-order polynomial as expected
values, we can use a Chi-squared measurement to quantify the
degree to which the observed Falcon 9 and Long March 3C
data match those of the lunar impactor. The Chi-squared value
for the Long March 3C R/B (0.83) is lower (better match) than
that of the Falcon 9 R/B (2.02), indicating that the Long March
3C R/B spectrum is a better match to the spectrum of the lunar
impactor, further supporting the hypothesis that the lunar
impactor was indeed the Long March 3C upper stage that
launched the Chang’e 5-T1 mission. We will refer to
WE0913A in the subsequent sections as the Chang’e 5-T1
R/B for clarity and conciseness.
Since all spectra involved in this analysis are relatively
featureless, it is worth discussing potential causes of the
difference in the spectral slopes seen in Figure 4 to better
understand the implications for the object’s origin. Potential
explanations for the differences between these spectra include
phase angle, space aging, and compositional differences
(Pearce et al. 2020; Battle et al. 2023). Comparison objects
were chosen that were launched at a similar time as the
proposed origins of WE0913A so that the space aging impacts
potentially would be similar and could largely be ignored. It is
possible that WE0913A received higher amounts of radiation in
deep space, and thus more space aging, compared to the
comparison R/B that remained in low-Earth orbit (Reitz 2008;
Restier-Verlet et al. 2021). Phase angle effects are harder to
address, since there is no singular library of space-aged
spacecraft material spectra at multiple phase angles to help
address how phase angle effects manifest for different
materials.
Additionally, we do not know the composition of the
comparison R/Bs, which makes all comparisons more
difficult. The major components of these R/B spectra,
however, will typically include a white thermal control paint
used on the exterior of the fuselage and the metal used for the
engines. The composition of the thermal control material
(TCM) used and its resistance to space aging may vary
drastically between the Falcon 9 R/B and the Long March 3C
R/B. Some white TCMs tend to yellow as they space age,
while others are resistant to the changes and maintain a white
appearance (Dever 1991; Bengtson et al. 2018; Goto et al.
2021). Although the exact paint is not known, it is plausible
that the paint used on the Long March 3C R/B tends to yellow
with exposure to the space environment, whereas the paint
used on the Falcon 9 does not. This would create the observed
difference in spectra between the two objects, despite similar
times on orbit.
5. Photometric Analysis
Light curve inversion is a topic that has been extensively
studied as it applies to asteroids and other natural bodies
(Kaasalainen et al. 2001; Kaasalainen & Torppa 2001;
Muinonen et al. 2020), the limitations of which, and general
ill-posed nature of the problem, are well understood (Calef
et al. 2006; Hinks et al. 2013); however, useful information can
still be learned from even ill-posed inverse problems. These
techniques have also been no less frequently applied toward
artificial objects (Linares et al. 2013, 2014, 2020; Linares &
Crassidis 2018). The focus has, however, almost exclusively
been for characterization of shape or object type. There have
been few attempts in the literature to apply these methods to a
(partially) known object to assess the performance of a model-
based approach with a comparison to real data. One such
example is by Campbell et al. (2022), where they use this
technique to look at cislunar object 2020 SO and recover
several dynamical and physical properties. Originally dis-
covered by Pan-STARRS1 and thought to be a natural object, it
was later shown that 2020 SO was actually the Centaur upper
stage from the Surveyor 2 mission. Another example (Santoni
et al. 2013) also looks at recovering properties other than
shape, but there are some clerical errors in the material
characteristics and thus only diffuse reflection is considered,
which is not sufficient for human-made objects with highly
specular components. Moreover, the techniques used are shape-
specific and significant rotational assumptions are made, so the
methods are not applicable in general.
The photometric measurements taken on 2022 January 20
and February 8 show a strong apparent periodicity in the light
curves of the Chang’e 5-T1 R/B that implies the object is
neither actively stabilized nor tumbling (in the complex sense),
but rather uniformly rotating (typically) about either its major
or minor principal axes of inertia. It is also typical for objects
rotating in this manner to then have a precession of this
rotational axis about the body’s angular momentum vector
(Benson et al. 2017). To quantify this observed periodicity, we
performed a four-parameter Fourier fit with least squares
minimization to the light curves taken just before each of two
sequential close Earth flybys. The results of these minimiza-
tions are in Figures 5(a) and (b).
Figures 5(a) and (b) show that the period of the light curve
changes between the two observing epochs, indicating a likely
change in the rotational period of the Chang’e 5-T1 R/B.
While the period of the light curve just before the second Earth
flyby is within the 2σ bounds of the first set of data, the reverse
is not true, indicating that this is likely a real change in the
rocket body’s rotational rate. A high-fidelity attitude simulation
and analysis may be able to lend insight into the impact of the
first Earth flyby on the Chang’e 5-T1 R/B rotational period,
but this is outside the scope of this work.
Using a predictive light curve model and Markov Chain
Monte Carlo (MCMC) sampling-based inversion using
astrometric and photometric data collected of the Chang’e
5-T1 R/B during both of its final close approaches with the
Earth before impacting the Moon, we recover likely
distributions for spin state and reflective properties. We
estimate nine parameters: primary body axis orientation (2),
angular velocity vector (3), diffusive/specular reflectivity
parameters (2), and surface anisotropic/roughness para-
meters (2).
4
The Planetary Science Journal, 4:217 (12pp), 2023 November Campbell et al.
5. 6. Bayesian Inversion Methodology
Bayesian inversion relies heavily on a statistical forward
model that can map arbitrary samples in your chosen parameter
space to realistic samples in the desired observation space. It is
important to strike a good compromise between model fidelity
and computational requirements for said model, since there are
harsh diminishing returns for increased fidelity with ill-posed
inverse problems.
In this research, our forward model takes as input an
arbitrary set of the nine parameters we want to estimate:
primary body axis orientation (2), angular velocity vector (3),
diffusive/specular reflectivity parameters (2), and surface
anisotropic/roughness parameters (2). With these parameters
and a previously selected shape, the model generates an
associated light curve. An in-depth discussion of the model
used is given in Campbell et al. (2022), and a summary of this
model is given in Section 6.1, while a description of our
Bayesian inversion methodology is given in Section 6.2.
6.1. Dynamical Model
The forward light curve model used in this research is
composed of three parts. There is the attitude model, which
handles the orientation of the body; the translational model,
which handles the location of the body; and the reflective
model, which handles the apparent brightness of the body. A
complete description of each of these sections is given in
Campbell et al. (2022), but the relevant equations are given
below without derivation.
6.1.1. Attitude Dynamics
We represent a body’s orientation as a unit quaternion (q̄)
with Euler parameters defined as a single rotation of angle j
measured in a counterclockwise direction about some well-
chosen unit vector ê. The expression of Euler parameters as a
unit quaternion is given in Equation (1), where q and q4 are the
vector and scalar parts of the quaternion, respectively:
*
⎡
⎣
⎢
⎢
⎢
⎛
⎝
⎞
⎠
⎛
⎝
⎞
⎠
⎤
⎦
⎥
⎥
⎥
⎡
⎣
⎤
⎦
j
j
= =
¯
ˆ
( )
q
e
q
q
sin
2
cos
2
. 1
4
The time derivative of the Euler parameters is given as:
⎡
⎣
⎢
⎤
⎦
⎥
=
+
-
¯
[ ] [˜]
( )
q
I q
q
w
q
1
2
, 2
T
4 3
where w is the angular velocity of the target in the body
reference frame, [I3] is the 3 × 3 identity matrix, and [˜]
q is the
vector cross-product matrix of the vector components of the
attitude quaternion. The time derivative of the angular velocity
is derived from the expression for total angular momentum and
is given as Euler’s equations of rotational motion in
Equation (3):
= - +
-
[ ] ( [ ˜][ ] ) ( )
w J w J w L . 3
1
During the Bayesian inversion process, we only consider a
short time history of observations, typically a few minutes, and
our target body is uncontrolled. As such, we assume no
appreciable external torques on the system and therefore treat
the motion as torque-free rotation during simulation (L = 0).
We also take the inertia tensor [J] to be constant and symmetric
in the body frame.
6.1.2. Orbital Dynamics
Before the Chang’e 5-T1 R/B crashed into the Moon on
2022 March 4, it had two close approaches with the Earth. The
second close approach in early February marked the last time
that the rocket body would be optically visible until its collision
with the Moon due to low solar elongation. Figure 6 shows the
geocentric orbit of the Chang’e 5-T1 R/B (black) in the Earth
mean equator mean equinox (MEME) J2000 reference frame
from 2022 January to 2022 March. The Moon’s orbit (gray) is
included as a reference.
Using a sixth-degree zonal gravity model (Equation (4)), the
trajectory of the Chang’e 5-T1 R/B can be precomputed before
the actual Bayesian inversion process to reduce computation
time. The requisite solar and observer positions at the
appropriate observation epochs can also be precomputed to
further improve efficiency. The short time window of
Figure 5. Light curves of Chang’e 5-T1 R/B from before each of two
sequential close Earth flybys taken from the same observatory. The light curves
have been period wrapped by fitting a four-parameter Fourier series to the data
via least squares minimization. As can be seen by the fitted period in each plot,
the observed rotation rate of Chang’e 5-T1 changed substantially between the
two flybys.
5
The Planetary Science Journal, 4:217 (12pp), 2023 November Campbell et al.
6. observations and relatively large distance to the Chang’e 5-T1
R/B mean that higher-order perturbations (such as tesseral or
sectoral harmonics) can be safely ignored during the trajectory
integration. Since this Bayesian inversion model is very
sensitive to attitude and reflective properties, any accuracy
benefits from higher-order perturbations (including SRP)
would be lost to the relative uncertainty of the inversion:
⎛
⎝
⎞
⎠
å
q j q
= Q
=
¥ +
( ) ( ) ( )
U r
r
A
, ,
1
cos . 4
l
l
l l
0
1
In Equation (4), U(r, θ, j) is the generalized zonal
gravitational potential of the Earth in spherical coordinates.
When considering only zonal effects, the j dependence
reduces to the degree-dependent constants Al and the θ
dependence reduces to the generalized (associated) Legendre
polynomials represented by q
Q ( )
cos
l .
6.1.3. Reflective Dynamics
The simulated data used in the estimation process are a
function of the location, shape, attitude, and reflective
properties of the target body (Chang’e 5-T1 R/B), as well as
the relative location of the Sun and observer (Linares et al.
2013). The simulated flux of the target at any instant in time
may be calculated by:
å
r
=
=
(ˆ · ˆ )(ˆ · ˆ )
( )
u n u n
F
A
d
. 5
i
N
i i sun i i
tgt
1
obs
2
The total flux received by a sensor from the target, Ftgt, may
be calculated as the sum of the flux from all facets of the target
body. In this case, the target body is represented as a digital
shape model described by many small triangular plates. A
diagram of the layout of an individual body facet is given in
Figure 7.
The BRDF contribution, ρi, from each facet is multiplied by
the area of each facet, Ai, projected in both the incident and
reflecting directions (given by the dot products of the unit
vectors in the Sun and observer directions with the surface
normal). The flux density decreases with d
1
2 , where d is the
observer–target distance (Hinks et al. 2013). According to the
modified Phong BRDF of Ashikmin & Shirley (2000), the total
reflectance for each facet is the sum of the specular, ρs,i, and
diffuse, ρd,i, components of the BRDF.
Modeling anisotropy in the desired material can be
controlled by the constants nu and nv, which influence the
specularity in each facet direction (see Figure 7 for the facet
basis definition and Campbell et al. 2022 for the specific form
of the BRDF used):
= - + ( )
m F K
2.5 log . 6
app 10 tgt
The total flux calculated may then be used as in Equation (6)
to calculate an estimated magnitude for the target object. The
constant K is any calibration that needs to be applied. This
calibration includes the zero-point offset for whichever spectral
band or filter you are working with, as well as any noise,
airmass extinction, or other effects being modeled that are not
included in Ftgt. In this paper, K needs only to be set to the
apparent solar magnitude in the appropriate filter for the
observations being processed. See Section 2 for details on the
filters used to collect observations.
6.2. Bayesian Estimation
In Bayesian statistics, the estimated posterior likelihood of
an event is determined by any observational data available and
any a priori assumptions about the system. Bayes’ rule allows
the estimation of the posterior conditional probability
distributions q
( ∣ ˜)
y
p with the aid of a probabilistic model that
associates all chosen parameters θ to observations ỹ via a
likelihood function (Aster et al. 2018):
ò
q
q q
q q q
=
( ∣ ˜)
( ˜∣ ) ( )
( ˜∣ ) ( )
( )
y
y
y
p
p p
p p d
. 7
The integral in the denominator of Equation (7) is just ( ˜)
y
p
and serves to normalize q
( ∣ ˜)
y
p so that it may be considered a
proper probability density function. Since direct computation of
the posterior distributions for an arbitrary choice of parameters
is challenging in a low-dimensional well-posed problem and
impossible in an ill-posed problem, we employ a method
known as MCMC sampling. This allows for the generation of
values from an estimated posterior distribution that fits the
observational data that you have and any a priori assumptions.
The Metropolis–Hastings (MH) sampling method for MCMC
simulations (Figure 8) is formulated so that the integral
Figure 6. Example of the Chang’e 5-T1 R/B orbit (black curve) showing close
approaches to the Earth between 2022 January until 2022 March, when it
collided with the Moon, shown in the geocentric MEME J2000 equatorial
reference frame. The Moon’s orbit (gray curve) and Earth’s location (black dot)
are included for reference.
Figure 7. Illustration of the reference basis and relevant observation vectors on
a shape model body facet.
6
The Planetary Science Journal, 4:217 (12pp), 2023 November Campbell et al.
7. constant in the denominator of Equation (7) cancels out during
the computation, allowing for the posterior distributions to be
accurately estimated.
This Markov chain method tests candidate samples with the
acceptance probability α to ensure that the Markov chain tends
toward areas of the parameter phase space with a higher
posterior estimate. This means that we may use a computa-
tional model as described in Sections 6.1.1–6.1.3 together with
an evaluation of the goodness of fit to the observed data (sum
of square residuals) to evaluate the likelihood of each MH-
generated sample from the parameter phase space, q
( ˜∣ )
y
p . We
may also supply an a priori estimate of the distributions to
which our parameters belong (p(θ)), which can be unin-
formative (uniform distribution) if we have no strong a priori
information.
While this method does produce a valid Markov chain, the
Markov property is not what allows successful estimation of
the posteriors. It is instead the fact that the estimated
distributions are improved with each sampling, thus converging
to the true distributions in the limit. Through continued
generation of this Markov chain, we may accurately estimate
the posterior distributions and only need to check the Markov
chain for convergence to ensure a valid estimate of the
posteriors.
7. Bayesian Analysis and Results
We ran a total of 100,000 iterations of our MCMC inversion
on the data collected with the Leo-20 telescope on 2022
January 20 (see Table 2). Our prior distributions for each
estimated parameter are partially uninformative, in that they are
not true (infinite) uniform distributions, but rather finite
distributions constrained by what is reasonable for both the
model and the physics of the problem. A summary of the prior
distributions for each parameter is given in Table 3. The
parameters α and β describe the primary body axis orientation
at the start of the observation epoch, w1−3 make the angular
velocity vector, and ρ and F0 control the relative diffusivity and
specularity of the body, respectively. Last, nu and nv control the
amount of anisotropy in each of the body facet directions (see
Figure 7).
As is the case for the majority of inverse problems of even
moderate complexity, this estimation problem is ill-posed, so
the results presented here represent a “best estimate.” That is,
due to the difficulty of sampling the complete surface of the
Chang’e 5-T1 R/B and the large amount of symmetry on the
rocket body, recovering the true states from this analysis is
unlikely. However, there exist metrics that can be used to
gauge the quality of the resulting estimates to provide
confidence in the results.
By design, the MCMC estimation process is prone to getting
“stuck” near the initial state estimation for some time. In order
to ensure that the parameter space is diversely sampled and that
we are not just sampling near our initial guess, we can look at
the chain of all parameters chosen during the inversion process.
This inertia-like property of MCMC algorithms is known as
“burn-in” and is important to be aware of when using
uninformative prior distributions. Looking at Figure 9, we
see that the sampling chain, after the initial burn-in is removed,
is highly varied and thus samples the parameter space
sufficiently well.
A quick quantitative way to evaluate the performance of the
MCMC sampling is by checking the acceptance rate. Samples
are only added to the posterior distribution estimate if they pass
a strict sampling gate test. Ideally, the acceptance rate should
be between 20% and 50% (Aster et al. 2018), and for our test
we had an acceptance rate of 38.18%. If the acceptance rate is
too low, then the estimation did not sufficiently sample the span
of the parameter space and will not accurately portray the
regions of highest posterior density (HPD). Conversely, if the
acceptance rate is too high, then it is likely that the algorithm
was stuck in a local extremum, and again we cannot be
confident that it sufficiently sampled the span of the parameter
space for globally valid results.
Figure 10 shows diagrams of the posterior distribution
estimates for each parameter, and Table 4 summarizes the
relevant parameters associated with the distributions for each
parameter.
An important distinction for the values provided in Table 4 is
the subtle difference between frequentist and Bayesian statistics
as they are given here. The frequentist statistics (mean and
standard deviation) are indicative of how often an event occurs,
while the Bayesian statistics (maximum a posteriori, or MAP,
and HPD) provide how likely it is that an event occurs, given
the data you have and your understanding of the system.
Looking at Figure 10 and comparing the values in Table 4, we
see that in this case, the estimated distributions for our
parameters can be well approximated by Gaussian
Figure 8. Brief overview of the MH algorithm for MCMC sampling.
Table 3
Partially Uninformative Prior Distribution for Each Parameter to Be Estimated
by the Bayesian Inversion
Variable Initial Value Min Max Distribution
α 0.1 0 2π U(0, 2π)
β 0.1 0 2π U(0, 2π)
w1 0.000 1 −10 10 U(− 10, 10)
w2 0.000 1 −10 10 U(− 10, 10)
w3 0.000 1 −10 10 U(− 10, 10)
ρ 0.01 0 1 U(0, 1)
F0 0.001 0 1 U(0, 1)
nu 10 −1000 1000 U(− 1000, 1000)
nv 10 −1000 1000 U(− 1000, 1000)
7
The Planetary Science Journal, 4:217 (12pp), 2023 November Campbell et al.
8. distributions. This is typically not the case in general, as the
estimated distributions from this process can be highly non-
Gaussian.
Estimating the complete distribution of each parameter not
only gives us insight into the global set of possible results, but
also allows sampling from each distribution to evaluate
behavior. In Figure 11, we took 1500 random samples from
each distribution and fed them back into our forward model.
Plotting the results and overlaying the raw Chang’e 5-T1 light
curve, we can see that the estimated parameters and forward
model fit the data very well. Note that the diagonal portion of
the plot between 17 and 21 minutes post-epoch is when there
were no data. Since it is possible for the “best estimate” to
converge on an incorrect or poorly fitted solution, evaluating
the results in this way offers a self-consistency check and
confidence that the estimated parameters well describe the data.
Based on the values in Table 4, it would appear that the
Chang’e 5-T1 R/B is very anisotropic (ratio of nu to nv). As a
reminder, these two parameters control the surface roughness and
the degree of anisotropic reflection in either the û or v̂ directions
for each body facet (see Figure 7). In Table 4, we see that the
estimated reflection is approximately 60 times greater in the û
direction than the v̂ direction. While it is quite likely that the
Chang’e 5-T1 R/B exhibits some anisotropy, this level is
unexpected. This result is likely due to the fact that the
observations used for this estimation spanned a short time
window, and so geometrically speaking, reflection in the
orthogonal axis may not have been as visible. This is further
supported by the rotation of the body being that of a very stable
spin, implying that body surfaces are observed from the same
geometry over short timescales as the body rotates. This is
opposed to a complex tumbling body, where the observer–target
body surface geometry may not be as periodic during rotation.
The strong periodic fits seen in Figures 5(a) and (b) imply
that the Chang’e 5-T1 R/B was in a very uniform spin during
the observations. This is reflected in our results by the
estimated angular velocity (Table 4). Note that it is almost
identically equal to a uniform spin about a single (the principal)
body axis. This is somewhat unexpected for a typical rocket
body, especially one that has been on-orbit for almost a decade
and has had multiple close encounters with the Earth and
Moon. In a typical (empty) rocket upper stage, the majority of
Figure 9. Time history of parameter sampling from the Bayesian inversion where the first 55,000 samples have been removed to avoid burn-in.
8
The Planetary Science Journal, 4:217 (12pp), 2023 November Campbell et al.
9. its mass is concentrated toward the engines, giving an
appreciable center-of-mass–center-of-figure (COM–COF) off-
set. This COM–COF offset can cause perturbations such as
SRP or gravity gradient torque to have a larger destabilizing
effect, giving rise to more complicated motion than just simple
spin (i.e., precession, nutation, complex tumble, etc.; Takahashi
et al. 2013), which is one possible explanation of the results
seen in Campbell et al. (2022). However, we see no such
evidence for perturbations in the spin of the Chang’e 5-T1
R/B, implying that the COM–COF offset may be smaller than
expected. This would be the case if there was additional mass at
the front of the rocket body, opposite the engines. This
additional mass at the front of the body would bring the COM
forward, closer to the COF, and thus reduce the effect of these
destabilizing perturbations.
The Chang’e 5-T1 R/B hosted a secondary payload (in
addition to the Chang’e 5-T1 spacecraft) that was comprised of
two instruments and dubbed the Manfred Memorial Moon
Mission by LuxSpace (Siddiqi 2018). This payload has a
published mass of 10–14 kg and was permanently affixed to the
front of the Chang’e 5-T1 R/B (Clark 2014; Moser et al. 2015;
Siddiqi 2018). It is unclear if, or how, the upper stage was
Figure 10. Final estimates of the posterior distributions of each parameter.
Table 4
Bayesian and Frequentist Statistics for the Estimated Posterior Distributions of
Each Parameter
Variable MAP 95% HPD Mean STD
α 0.002 [0.0000, 0.0204] 0.008 0.008
β 0.287 [0.2473, 0.3008] 0.282 0.013
w1 0.000 [–0.0001, 0.0001] 0.000 0.000
w2 0.000 [–0.0001, 0.0001] 0.000 0.000
w3 0.017 [0.0174, 0.0176] 0.017 0.000
ρ 0.004 [0.0025 0.0054] 0.004 0.001
F0 0.037 [0.0276, 0.0500] 0.037 0.006
nu 30.203 [30.1695, 30.2221] 30.197 0.013
nv 0.497 [0.3995, 0.4999] 0.465 0.030
9
The Planetary Science Journal, 4:217 (12pp), 2023 November Campbell et al.
10. modified to accommodate this payload vis-à-vis structural
support or additional equipment that might affect the final mass
distribution, but the published mass of these instruments is only
up to 14 kg, which is not enough to appreciably move the COM
forward from the large mass of the dual-mounted YF-75
engines on its own (Cen et al. 2011).
One way to verify the hypothesis that there was additional
mass in the front of the rocket body is to look for the crater that
would result from the Chang’e 5-T1 R/B impact with the lunar
surface. Craters have been found previously on the lunar
surface following the impacts of rocket bodies from the Apollo
missions. However, searching for a small impact crater among
the plethora of craters on the Moon is challenging at best.
Hence, predicting the precise location of the impact is critical
for any hope of verifying this hypothesis.
8. Lunar Impact Location Prediction
The ephemeris of the Chang’e 5-T1 R/B was estimated by
fitting orbits to right ascension and declination measurements
reported by ground-based optical telescopes. A total of 508
observations were reported by 26 different stations from 2021
July 8 to 2022 February 10.
It was evident early on as observations accumulated that
nongravitational accelerations had a significant effect on the
trajectory. We used a typical radiation-based nongravitational
perturbation model, where the accelerations radial, transverse,
and normal to the heliocentric orbit are given in Equation (8)
(Marsden et al. 1973):
⎛
⎝
⎞
⎠
⎛
⎝
⎞
⎠
⎛
⎝
⎞
⎠
=
=
= ( )
a A
r
a A
r
a A
r
1 au
1 au
1 au
. 8
R
T
N
1
2
2
2
3
2
A1, A2, and A3 are estimable parameters during the orbit
fitting process. This model substantially improved fits over a
purely ballistic trajectory, though it was nevertheless an
imperfect match to the observations. This may have been due
to outgassing from the rocket body or caused by a complex
shape or rotation state. To limit the prediction errors stemming
from the nongravitational perturbation model, we substantially
shortened the fitted data arc, eventually settling on two orbit
solution cases: one long and one short arc. After outlier
rejection, the long arc included 162 observations spanning 57
days from 2021 December 15 to 2022 February 10. The short
arc was of approximately half the duration, with 154
observations covering 28 days from 2022 January 13 to 2022
February 10. We also tested a simplified nongravitational
acceleration model that included only the A1 term (represent-
ing SRP).
Figure 12 shows the predictions for these four test solutions,
along with their associated 3σ uncertainty ellipses. The limited
overlap among these solutions emphasizes the difficulty in
nongravitational perturbation modeling. As a result, we elected
to use the three-parameter nongravitational perturbation model
in light of the optimistic uncertainties demonstrated by the one-
parameter model. Thus, solution S23 was the final prediction,
for which we document the orbital parameters in Table 5. The
estimated radial term of the nongravitational acceleration
Figure 11. The 97.5% HPD region (light gray) in light curve phase space as
estimated by 1500 random samples from the posterior estimates generated with
data taken from the Leo-20 telescope on 2022 January 20. The observed light
curve of the Chang’e 5-T1 R/B has been overlaid in black. The diagonal
portion of the light curve from approximately 17–21 minutes post-epoch is a
period with no data.
Figure 12. Impact prediction for several trajectory solutions as described in the
text. The final delivered prediction was for solution “S23.”
Table 5
Estimated Nongravitational and Osculating Orbital Parameters for Solution S23
Giving the Predicted Lunar Impact Location
Parameter Value 1σ Uncertainty
Epoch (TDB) 2022 Feb 05 00:00:00 N/A
a (km) 321774.936 0.317
e 0.912983869 2.96E-7
i (deg) 31.4782088 4.52E-5
ω (deg) 151.4305753 8.02E-5
Ω (deg) 14.3057527 5.80E-5
ν (deg) 196.0218047 2.69E-5
A1 (km s−2
) 42.597E-12 1.505E-12
A2 (km s−2
) –1.808E-12 0.936E-12
A3 (km s−2
) 2.079E-12 0.523E-12
Note. Osculating elements are given in the Earth-centered MEME J2000
equatorial reference frame.
10
The Planetary Science Journal, 4:217 (12pp), 2023 November Campbell et al.
11. model (A1) corresponds to an area-to-mass ratio of approxi-
mately 0.01 m2
kg−1
. Based on the scatter among the
predictions seen in Figure 12, we increased the impact location
prediction uncertainties to ±4 km in longitude and +1/−3 km
in latitude. Figure 12 also depicts the reported impact location
from the Lunar Reconnaissance Orbiter (LRO), which indicates
that the prediction was in error by 7.0 km in longitude and
2.7 km in latitude.
9. Impact Crater Morphology
Given the relatively high degree of certainty of the impact
location of the Chang’e 5-T1 R/B on the lunar surface, there
was a decent chance that it could be found. Fortunately, based
on our predicted impact location, the LRO was able to find and
image the impact site using the LRO narrow-angle camera
(Robinson et al. 2010). Comparing LRO images of the impact
site both pre- and post-impact gave rise to new questions.
Instead of one round, or elongated, crater, such as what was
seen when the Apollo Saturn IV-B rockets were sent into the
Moon, there are two distinct craters side by side that were made
by the Chang’e 5-T1 R/B. Figures 13(a) and (b) show the
relevant section of lunar terrain both pre- and post-impact.
The double crater (pictured in Figure 14) formed by the
Chang’e 5-T1 R/B impact with the Moon was unexpected
when considering what has been seen from other rocket body
impacts with the Moon (Figure 15). Four Apollo Saturn IV-B
rocket bodies were crashed into the Moon and created
somewhat elongated craters, though they had shallow impact
angles. The Chang’e 5-T1 R/B impact with the surface came
from roughly 15° off vertical (75° impact angle), so that is not
the case here. Moreover, the appearance of the craters is not of
one elongated shape made from a single impact, but rather two
distinct craters of similar size formed from separate impacts of
similar energy. This supports our hypothesis that there was
additional mass on the front end of the Chang’e 5-T1 R/B,
creating the second crater during impact.
As noted earlier, there were additional instruments affixed to
the front of the Chang’e 5-T1 R/B that were to stay attached to
the rocket body post spacecraft detachment. The published
mass of these instruments is only up to 14 kg (Clark 2014;
Moser et al. 2015; Siddiqi 2018), which is not enough alone to
account for the double crater seen post-impact.
Figure 13. These are approximately 500 m sections of the lunar terrain captured by LRO showing the Chang’e 5-T1 R/B impact site and double crater that was
formed. Before is from LRO image M1400727806L and after is from M1407760984R (NASA/GSFC/Arizona State University). Other visual differences in the
terrain are caused primarily by the difference in the angle of solar incidence, which was approximately 28° during the before image and 59° in the after image. This
section of terrain is near the Hertzsprung crater on the lunar far side.
Figure 14. Zoomed-in view of the double crater formed by the Chang’e 5-T1
R/B impact with the Moon. Context images in Figures 13(a) and (b). Data
from LRO image M1407760984R (NASA/GSFC/Arizona State University).
11
The Planetary Science Journal, 4:217 (12pp), 2023 November Campbell et al.
12. 10. Conclusions
In late 2021, it was discovered that an object (WE0913A)
would impact the Moon in 2022 March after several close flybys
of the Earth and the Moon over the coming months. The true
identity of this object was up for debate, with two possibilities:
(1) the Falcon 9 R/B from the DSCOVR mission; and (2) the
Long March 3C R/B from the Chang’e 5-T1 mission. Our
trajectory and spectroscopic analyses using ground-based
telescope observations show conclusively that WE0913A is in
fact the Chang’e 5-T1 R/B. Analysis of photometric light
curves gave a spin period of 185.221 ± 6.540 s at a 1σ
confidence level in the light curve of the Chang’e 5-T1 R/B just
before the first close Earth flyby and a period of
177.754 ± 0.779 s at a 1σ confidence level just before the
second close Earth flyby. Using MCMC sampling and a
predictive light curve simulation based on an anisotropic Phong
reflection model, we estimate both physical and dynamical
properties of the Chang’e 5-T1 R/B at the start of an
observation epoch. The Bayesian analysis implies that there
was an additional mass at the front of the R/B that stabilizes the
spin characteristics. We used the observations acquired to
pinpoint the location of the impact on the lunar surface, which
enabled the LRO to find and photograph the crater site. Our
predicted location is within 7 km in longitude and 2.7 km in
latitude from where the crater was eventually found by LRO.
Comparing the pre- and post-impact images of the location
shows two distinct craters side by side that were made by the
Chang’e 5-T1 R/B. The double crater supports the hypothesis
that there was additional mass at the front end of the rocket
body, opposite the engines, in excess of the published mass of
the secondary permanently affixed payload.
Acknowledgments
This work is supported in part by the state of Arizona
Technology Research Initiative Fund (TRIF) grant (PI: Prof.
Vishnu Reddy). A portion of this work was conducted at the Jet
Propulsion Laboratory, California Institute of Technology, under a
contract with the National Aeronautics and Space Administration.
ORCID iDs
Tanner Campbell https:/
/orcid.org/0000-0001-9418-1663
Adam Battle https:/
/orcid.org/0000-0002-4412-5732
Steven R. Chesley https:/
/orcid.org/0000-0003-3240-6497
Davide Farnocchia https:/
/orcid.org/0000-0003-0774-884X
Neil Pearson https:/
/orcid.org/0000-0002-0183-1581
Vishnu Reddy https:/
/orcid.org/0000-0002-7743-3491
Roberto Furfaro https:/
/orcid.org/0000-0001-6076-8992
References
Andrae, R., Fouesneau, M., Creevey, O., et al. 2018, A&A, 616, A8
Ashikmin, M., & Shirley, P. 2000, Journal of Graphics Tools, 5, 25
Aster, R., Borchers, B., & Thurber, C. 2018, Parameter Estimation and Inverse
Problems (Amsterdam: Elsevier)
Battle, A., Reddy, V., Furfaro, R., & Campbell, T. 2023, AcAau, submitted
Battle, A., Reddy, V., Sanchez, J., et al. 2022, PSJ, 3, 226
Bengtson, M., Maxwell, J., Hoffmann, R., et al. 2018, in The Advanced Maui Optical
and Space Surveillance Technologies Conf., ed. S. Ryan (Kihei, HI: AMOS),33
Benson, C., Scheeres, D., & Moskovitz, N. 2017, in 7th European Conf. on
Space Debris (Paris: ESA), 117
Bertin, E. 2006, in ASP Conf. Ser. 351, Astronomical Data Analysis Software
and Systems XV, ed. C. Gabriel et al. (San Francisco, CA: ASP), 112
Bertin, E., & Arnouts, S. 1996, A&AS, 117, 393
Calef, B., Africano, J., Birge, B., Hall, D., & Kervin, P. 2006, Proc. SPIE,
6307, 63070E
Campbell, T., Furfaro, R., Reddy, V., et al. 2022, JAnSc, 69, 95
Cen, Z., Gao, R., & Shen, L. 2011, LM-3A Series Launch Vehicle User's
Manual (Beijing: China Academy of Launch Vehicle Technology)
Christensen, E., Africano, B., Farneth, G., et al. 2018, AAS/DPS Meeting, 50,
310.10
Clark, S. 2018, Spaceflight Now, https://spaceflightnow.com/2014/10/25/
first-commercial-mission-to-the-moon-launched-from-china/
Dever, J. 1991, Low Earth Orbital Atomic Oxygen and Ultraviolet Radiation
Effects on Polymers NASA-TM-103711, NASA Lewis Research Center,
https://ntrs.nasa.gov/citations/19910009981
Gaia Collaboration 2018, Gaia Data Release 2 (Gaia DR2), ESA, doi:10.5270/
esa-ycsawu7
Goto, A., Umeda, K., Yukumatsu, K., & Kimoto, Y. 2021, CEAS, 13, 415
Hinks, J., Linares, R., & Crassidis, J. 2013, in AIAA Guidance, Navigation,
and Control (GNC) Conf. (Reston, VA: AIAA), 5005
Jorgensen, K., Culp, R., & Clark, R. 2001, in Proc. of the Third European
Conference on Space Debris (Noordwijk: ESA), 385
Kaasalainen, M., & Torppa, J. 2001, Icar, 153, 24
Kaasalainen, M., Torppa, J., & Muinonen, K. 2001, Icar, 153, 37
Linares, R., & Crassidis, J. 2018, JGCD, 41, 47
Linares, R., Crassidis, J., Wetterer, C., Hill, K., & Jah, M. 2013, Astrometric
and Photometric Data Fusion for Mass and Surface Material Estimation
using Refined Bidirectional Reflectance Distribution Functions-Solar
Radiation Pressure Model Tech. Rep. ADA593458, Pacific Defense
Solutions LLC, https://apps.dtic.mil/sti/tr/pdf/ADA593458.pdf
Linares, R., Furfaro, R., & Reddy, V. 2020, JAnSc, 67, 1063
Linares, R., Jah, M., Crassidis, J., & Nebelecky, C. 2014, JGCD, 37, 13
Marsden, B., Sekanina, Z., & Yeomans, D. 1973, AJ, 78, 211
Moser, H., Ruy, G., Schwarzenbarth, K., et al. 2015, in 29th Annual AIAA/
USU Conf. on Small Satellites, SSC15–V–7, https://digitalcommons.usu.
edu/cgi/viewcontent.cgi?article=3200&context=smallsat
Muinonen, K., Torppa, J., Wang, X., Cellino, A., & Penttilä, A. 2020, A&A,
642, A138
Pearce, E., Weiner, B., & Krantz, H. 2020, JSSE, 7, 376
Reitz, G. 2008, Zeitschrift für Medizinische Physik, 18, 233
Restier-Verlet, J., El-Nachef, L., Ferlazzo, M., et al. 2021, International Journal
of Molecular Sciences, 22, 3739
Robinson, M., Brylow, S., Tschimmel, M., et al. 2010, SSRv, 150, 81
Santoni, F., Cordelli, E., & Piergentili, F. 2013, JSpRo, 50, 701
Siddiqi, A. 2018, Beyond Earth: A Chronicle of Deep Space Exploration,
1958–2016, Vol. 4041 (Washington, DC: NASA)
Takahashi, Y., Busch, M., & Scheeres, D. 2013, AJ, 146, 95
Wenbin, W. 2022, in Foreign Ministry Spokesperson Wang Wenbin's Regular
Press Conference, February 21, 2022 (Ministry of Foreign Affairs of the
People’s Republic of China), https://www.fmprc.gov.cn/mfa_eng/xwfw_
665399/s2510_665401/202202/t20220221_10644075.html
Figure 15. Images of craters formed by the intentional impacts of the Apollo Saturn IV stages with the lunar surface. Note the round or elongated shapes, but all being
explicitly one crater. (NASA/GSFC/Arizona State University).
12
The Planetary Science Journal, 4:217 (12pp), 2023 November Campbell et al.